5.1 DSM Preprocessing 
During preprocessing the shape of every single tooth is 
determined. This is done semiautomatically. In a first 
step the center of every tooth is pointed manually in one 
image. In this step additional data (tooth number, ..) is 
set. Size and shape of the tooth is determined in a 
second step automatically. The gradient of the image is 
calculated in radial direction from the center pointed 
manually. If the gradient is larger than a certain 
threshold, if the direction of the gradient is rectangular to 
the radial direction within a certain range of tolerance and 
if the greyvalue of the pixel is smaller than a certain 
threshold the analysed pixel is classified as a point that 
belongs to the border of the tooth. All threshold values 
used are defined automatically by analysis of the 
greyvalues of the teeth and the gums. The shape of the 
tooth derived by this method is checked for gross errors 
and smoothed. In a final step the shape is approximated 
by an ellipse. All pixels inside the ellipse are classified as 
the image points that belong to one single tooth. 
  
Figure 8: Shape of teeth extracted automatically 
5.2 DSM Generation 
Automatic DSM generation is done by geometric 
constrained template matching (Gruen, Baltsavias 1988) 
and (Baltsavias 1992). To apply this algorithm images 
must be oriented. This is done by measuring the control 
points on the mirror in both images. After orientation and 
preprocessing a DSM for every single tooth is measured 
consisting of up to 900 surface points for a molar tooth. 
As usual for matching algorithms three problems appear: 
firstly the choice of a proper template size, secondly the 
derivation of approximations for the geometric 
parameters of the matching algorithms and finally the 
choice of a proper iteration criterion. The template size 
depends on the quality of the images. Molar teeth show 
more structure than front teeth and may therefore be 
matched with smaller templates (~15x15 pixels). Front 
teeth show, due to their smoothness, lower signals and 
have to be matched with larger templates of 
approximately 21x21 pixel size. (See results of tests 
reported in chapter 7) 
The derivation of approximations can be done in two 
different ways, that can be combined for one solution. 
The relative configuration of mirror and teeth is similar 
for any image acquisition. Assuming that all teeth are 
placed in a plane the imaging ray of one image can be 
intersected with this plane. This leads to 3D 
approximation coordinates that can be projected into the 
other image. This method can be combined with a 
manual measurement of the parallaxes by applying the 
preprocessing algorithm for shape and size extraction of 
all teeth in both images. This method leads to the best 
approximations for the DSM generation process. 
DSM generation using only the natural structure of the 
teeth is strongly influenced by reflections. Their bad 
influence on the measurement has to be eliminated. This 
is done by a very simple but also very efficient method. 
The influence of control points that overlay parts of teeth 
in the images can be eliminated too. Analysing the 
average greyvalue of all teeth during DSM preprocessing 
a threshold is defined. Any pixel that has a greyvalue 
larger than the threshold is skipped and eliminated from 
the estimation process. This method can be dangerous if 
too many pixels are skipped. In this case the surface 
point is eliminated from the DSM. 
Finally a proper iteration criterion has to be chosen. This 
is as problematic as for unconstrained LSTM. The main 
problem is the oscillations of transformation parameters. 
The problem is solved the same way as it is done for 
LSTM (Beyer 1992). In addition non determinable 
parameters have to be detected and eliminated from the 
estimation process. Strongly correlating parameters have 
to be excluded too. 
  
Figure 9: Result of DSM generation 
  
Figure 10: Result of postprocessing 
250 
International Archives of Photogrammetry and Remote Sensing. Vol. XXX, Part B5. Vienna 1996 
5.3 DSM Pc 
Since DSM 
structure of 
strong refle 
many mism 
quality of th 
illumination 
problem ev 
gross errors 
point of the 
a distance - 
threshold is 
and is more 
tooth corre 
points that 
interpolatior 
interpolated 
dimensional 
interpolated 
on the sam 
skipped anc 
step the DS 
algorithm wi 
6. DERIV/ 
The result c 
of all teeth 
space is de 
points on 
reference cc 
this coordin 
be done b 
decide wha 
the photogr 
Much more 
reference sj 
shall not mc 
reference sj 
orthodontist 
position anc 
object spac 
belonging t 
defined by 
shall move : 
the other h 
whether the 
not move ai 
of deform: 
geodesy. 
But the ma 
orientation « 
be defined c 
have to be c 
by a three-c 
algorithm i 
template m 
difference, | 
the third dir 
The transfo 
and 3 rotati 
algorithm is 
described b